Research

Research.

Two papers accepted at IEEE conferences in 2026.

IEEE ANTS 2026 · IIT Roorkee · 17–20 Dec 2026Accepted

LightDep: Knowledge-Distilled Multimodal Depression Screening for Mobile and Edge Devices

Devansh Pandey*, Prateek Goyal*, Dr. Amit Vishwakarma¹, Dr. Arijit Roy², Dr. Om Jee Pandey³

¹IIITDM Jabalpur·²IIT Patna·³IIT BHU

IEEE International Conference on Advanced Networks and Telecommunications Systems (ANTS)

The paper

A small model that screens for depression from face and voice features and runs on the device, even inside a browser tab. It learns from bigger teacher models (knowledge distillation) and is then quantized to 3.96 MB.

My part

Joint first author with Prateek Goyal. I built and trained the model, ran the distillation experiments and the significance tests, and did the INT8 ONNX export.From my summer research internship at IIT BHU.

Results

Distillation only helps when the teacher is a stronger, well-calibrated screener: +10.4 recall on D-Vlog (p = 0.045), confirmed over 10 seeds on a second GPU. No significant gain on LMVD.
3.96 MB INT8 model, 28× smaller than its teacher, 96 ms per inference on one CPU thread, 75.9 binary F1 on D-Vlog.

Full project write-up →Code ↗* equal contribution

IEEE TENCON 2026Accepted

Structure-Aware Drug–Target Interaction Prediction via Pretrained Language Models and Cross-Attention Fusion

Srajal Tiwari, Dr. Dolly Sharma, Prateek Goyal, Devansh Pandey, Anamika Pal, Abhinav Bachchas

IEEE Region 10 Conference (TENCON)

The paper

BioX-DTI predicts whether a drug will interact with a protein. Proteins are encoded with ESM-2 over AlphaFold contact graphs, drugs with a molecular graph plus ChemBERTa-2, and the two are fused with cross-attention.

My part

Co-author. I built the fusion module: 4-head cross-attention where each protein residue attends over the drug's atoms, followed by masked pooling.

Results

0.872 accuracy / 0.746 MCC on the Yamanishi 1:1 benchmark.
0.819 F1 on proteins the model never saw in training (cold-target split).

Code ↗